跳转至

Historical version

This page records design, operation, or validation facts from an earlier release. It is not current V7.17.3 guidance. Start from the documentation hub for current instructions.

Codex Cross-Project Long-Term Engineering Assistant V6.3 Validation Report

Subject

  • Version: 6.3.0
  • Target host: native Windows Codex CLI 0.150.1
  • Recommended form: account-level Plugin
  • Upgrade baselines: V6.1.0 and V6.2.0
  • Event contract: TaskOutcomeEvent 2.0
  • Automatic subagent ceiling: gpt-5.6-terra + high
  • Execution authorization: NONE

In-Package Validation

Item Result Evidence Entry
JSON/TOML parsing and Python compilation PASS scripts/validate-v63.py
Neutral language and structural semantics PASS scripts/semantic-lint.py
Skill routing cases PASS, 35 scripts/routing-eval.py validate
Package unit and regression tests PASS, 53/53 tests/test_*.py
Shared runtime regressions PASS, 6/6 runtime/tests/test_*.py
Installation transactions and fault injection PASS, targeted tests/test_package_manager_security.py
Lifecycle and self-observation quality PASS, targeted tests/test_v60_deterministic_observation.py
Deterministic ZIP and release-evidence contract PASS, targeted tests/test_v63_release_delivery.py
Evolution Proposal safety boundary PASS, targeted tests/test_v51_controlled_evolution.py

Every PASS in validate-v63.py derives from commands and suites executed in this run. Unexecuted real-host items remain NOT_EXECUTED; constants do not substitute for evidence.

Isolated Real-Machine Upgrade Loop

Completed with real Codex CLI 0.150.1 under a short isolated account path:

V6.1.0 installed/enabled
  -> V6.3.0 install + verify + Plugin list readback
  -> V6.3 uninstall
  -> V6.1.0 installed/enabled restored

Result: PASS.

  • V6.3 readback: installed=true, enabled=true, version=6.3.0.
  • After uninstall: installed=true, enabled=true, version=6.1.0.
  • No active transaction journal remained after V6.3 commit.
  • The isolated upgrade backup was retained.
  • The real account CODEX_HOME was not written during this isolated test.

The first isolated attempt used a deep workspace path and was blocked by V6.1's legacy long-path limitation before V6.3 wrote anything. Retesting with a short isolated path passed.

Independent Review

Preimplementation review identified gates for release evidence, real lifecycle, reproducible ZIP, observation completeness, and Reviewer attribution. Two logically read-only specialist Reviewers covered state/concurrency and test/delivery after implementation.

Major resolved findings:

  • per-target journal ownership for install, repository install, and uninstall;
  • persistence of Plugin and Marketplace state before uninstall unregistration;
  • complete Marketplace staging with expected target hash recorded before replacement;
  • lifecycle aggregation by session_id + task_id, with cross-session task conflicts entering the evidence gate;
  • duplicate event IDs no longer hiding later chain damage;
  • stale-lock logic no longer deleting a lock held by a live process based only on age;
  • validation counts and PASS states derived from actual output;
  • V6.3 release and audit files included in the package;
  • release evidence covers content tampering, illegal paths, correct/incorrect HMAC keys, and cross-parent lifecycle ordering errors.

Review was logically read-only. TaskOutcomeEvent readback reported gpt-5.6-luna; actual reasoning effort was not supplied by the host and remained unverified.

Formal Host Acceptance

Formal account acceptance on native Windows Codex CLI 0.150.1: PASS.

  • V6.1.0 -> V6.3.0 Plugin transaction upgrade succeeded.
  • codex plugin list --json readback: installed=true, enabled=true, version=6.3.0.
  • Ten Skills, seven Reviewers, and six Hooks were discoverable; Reviewer TOML did not fix a model.
  • Main configuration hash was unchanged and every preupgrade account Skill remained.
  • Historical project-context file count did not decrease and the upgrade backup remained.
  • No active transaction marker remained, and targets were not Junctions, Reparse Points, or symlinks.
  • A new read-only acceptance task started a cp_review_test_delivery Reviewer.
  • Diagnostic observation reported gpt-5.6-luna; the main-agent model configuration did not change.
  • One session produced TURN_OPENED, SUBAGENT_STARTED, SUBAGENT_STOPPED, TASK_COMPLETED, and SESSION_ENDED.
  • Five TaskOutcomeEvent 2.0 records had a continuous hash chain and matching project_id + repo_fingerprint.
  • SessionEnd completed within the three-second Hook limit; Windows used cp_hook.cmd without extra python3.exe.

The formal ZIP hash, byte-identical double build, hashes of the host evidence, and tamper verification are bound by an external machine proof. Without a valid machine proof, the artifact must not be called verified.

Security Boundaries

  • execution_authorization=NONE.
  • No automatic modification of Skills, Reviewers, main-agent models, or business code.
  • No automatic acceptance or implementation of Evolution Proposals.
  • No automatic commit, push, deployment, restart, or production operation.
  • Project aggregation validates both project_id + repo_fingerprint.
  • Historical Events, Snapshots, Assessments, Proposals, and upgrade backups survive ordinary upgrade and uninstall.